HypoLoop: Revising Weight-Editing Plans from Scientific Feedback
Organizations: Hong Kong Institute of Science and Innovation · Jinan University · University of Macau · Sichuan University ·
Abstract
Scientific experiments continually reveal prediction errors. Effective updates depend on which parameters change and which errors drive the correction. We introduce HypoLoop, a family of native weight-editing methods that connects these choices to experimental feedback. Fixed, numerically adaptive, and LLM-guided policies produce editing plans; numerical solvers compute weight patches, and internal validation selects the deployed predictor. Successive observations and validation records support revisions to editable locations and fitting rules. Experiments cover 22 scientific tasks and both DeepSeek and Claude. On Burgers, reallocating a matched parameter budget reduces error by 58.80% and 47.72% relative to fixed locations. On DIODE, positive-residual weighting reduces absolute relative error (AbsRel) by 13.35% in internal cross-validation, with the same 33 editable coordinates. Revising the complete plan across rounds lowers final AbsRel by 10.59% relative to keeping the initial plan fixed. SpaceNet7 and Make3D connect accepted edits to improved spatial predictions. These results show how experimental evidence can shape the update problem itself. Code: https://github.com/limengran98/HypoLoop.
Figures & tables
| DATASETS AND MODELS | BASE + FEEDBACK CONTROLS | HYPOLOOP FAMILY (OURS) | |||||||||||||
| Domain | ID / Dataset | Base model | Metric | Base | Adam [ 20 ] | Memory [ 35 ] | Calib [ 14 ] | HypoLoop | HL-Auto | HL-LLM | HL-Fixed | HL-Once | HL-Local | HL-Global | Best ∗ |
| Relative error reduction (%) | |||||||||||||||
| Biology | BIO01 IFN KO [ 12 ] | GEARS [ 33 ] | MSE | -50.2/-54.2 | -11.8/-15.8 | 0.0/0.0 | -48.1/-52.1 | +3.3 /-0.1 | +3.1 /0.0 | +2.8/ +0.1 | +2.8/ +0.1 | +0.5/-2.6 | -11.8/-15.8 | -9.1/-16.0 | +3.3/+0.1 |
| CHEM01 ESOL [ 9 ] | Uni-Mol [ 53 ] | RMSE | +4.7 / +4.7 | -13.2/-14.4 | -5.6/-5.6 | 0.0/ 0.0 | +11.5 /-21.9 | -18.0/-21.9 | -14.9/-27.9 | -18.0/-21.9 | +4.7 / +4.7 | -7.6/-15.5 | -15.0/-7.6 | +11.5/+4.7 | |
| Chemistry | CHEM02 Lipophilicity [ 47 ] | Uni-Mol [ 53 ] | RMSE | 0.0 /0.0 | -6.7/-6.7 | -5.9/-5.9 | 0.0 /0.0 | -21.6/-21.6 | +1.8 / +1.8 | -2.0/ +0.8 | -2.9/-2.9 | 0.0 /0.0 | -9.2/-7.2 | -16.1/-13.3 | +1.8/+1.8 |
| MAT01 Exfoliation [ 11 ] | CGCNN [ 48 ] | MAE | -4.4/-4.4 | 0.0/0.0 | -34.8/-34.8 | -0.4/-0.4 | -2.3/-2.3 | +0.6 / +0.6 | +0.5 / +0.3 | -0.3/-0.7 | -4.4/-4.4 | 0.0/0.0 | -3.5/0.0 | +0.6/+0.6 | |
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
| DATASETS AND METRICS | BASE + FEEDBACK CONTROLS | HYPOLOOP FAMILY (OURS) | |||||||||||
| Domain | ID / Dataset | Metric | Base | Adam | Memory | Calib | HypoLoop | HL- Auto | HL- LLM | HL- Fixed | HL- Once | HL- Local | HL- Global |
| Error metrics | |||||||||||||
| Biology | BIO01 IFN KO | MSE | 0.037298 0.037298 | 0.027770 0.028005 | 0.024837 0.024191 | 0.036793 0.036793 | 0.024028 0.024217 | 0.024058 0.024183 | 0.024134 0.024176 | 0.024134 0.024177 | 0.024711 0.024826 | 0.027773 0.028008 | 0.027104 0.028058 |
| Chemistry | CHEM01 ESOL | RMSE | 1.0246 1.0246 | 1.2172 1.2299 | 1.1348 1.1348 | 1.0749 1.0749 | 0.9518 1.3104 | 1.2688 1.3104 | 1.2347 1.3753 | 1.2688 1.3104 | 1.0246 1.0246 | 1.1565 1.2418 | 1.2367 1.1565 |
| CHEM02 Lipophilicity | RMSE | 0.9308 0.9308 | 0.9933 0.9933 | 0.9853 0.9853 | 0.9308 0.9308 | 1.1322 1.1322 | 0.9137 0.9137 | 0.9497 0.9229 | 0.9580 0.9580 | 0.9308 0.9308 | 1.0168 0.9981 | 1.0803 1.0549 | |
| Materials | MAT01 Exfoliation | MAE | 46.261 46.261 | 44.293 44.293 | 59.716 59.716 | 44.460 44.460 | 45.296 45.296 | 44.034 44.034 | 44.067 44.147 | 44.441 44.601 | 46.261 46.261 | 44.299 44.299 | 45.823 44.315 |
Model Editing
Model editing changes targeted facts or behaviors in trained models while aiming to preserve their other capabilities.
See all 71 papers